Tesla’s recently introduced 4680 cylindrical battery cell has higher energy density and power output than the currently available battery cells. Owing to its reduced surface-area-to-volume ratio that can lead to failure due to temperate rise, a thermo-mechanical crash analysis of the battery cell is crucial. The analysis can help to understand battery cell’s behavior under different loading conditions by varying temperature, displacement and strain rate. Moreover, a quick computation of the mechanical strength (force) of the battery cell using a data-driven machine learning approach can reduce the computational burden of re-analysis of the battery cells.
The methodology includes an integrated framework of thermo-mechanical crash analysis using finite element (FE) analysis and mechanical strength prediction using automated neural network search (ANS) modeling. In the first part, an FE model is developed by adopting the homogeneous material model for a jelly roll of the battery cell to avoid computationally intensive analysis of multiple layers consisting of anode, cathode, separator and electrolyte. The material model can represent the overall behavior of the jelly roll structure under specific loading conditions, making it reasonably accurate for applications such as crash analysis and structural response prediction. The response of varying temperature, displacement and strain rate on the mechanical strength of the 4680 cylindrical battery cell is stored and analyzed. In the second part, ANS modeling is adopted, in which different network models consisting of multi-layer perceptron (MLP), artificial neural network (ANN) and radial basis function (RBF) are used. These models are tested on the dataset developed through FE analysis in the first part by considering the different numbers of neurons in the hidden layer, training algorithms and activation functions.
Results of the thermo-mechanical analysis suggest that the casing of the battery cell experiences approximately 14 times higher stress than the jelly roll in order to protect it from damage. Moreover, the mechanical strength is majorly affected by displacement compared to temperature and strain rate variation. ANS modeling suggests that MLP having 29 neurons in the hidden layer with a logistic activation function shows the best performance based on correlation coefficient (R2) and mean square error (MSE) indicators. This model demonstrates outstanding performance in predicting mechanical strength, achieving an R2 value of 0.9997 and MSE of 0.0686.
The novelty of this paper lies in performing the thermo-mechanical crash analysis of Tesla’s 4680 cylindrical battery cell using FE analysis by varying the inputs, such as temperature, displacement and strain rate, to observe their effect on the cell’s mechanical strength. Moreover, the ANS model is developed based on FE analysis data for different values of input parameters to predict mechanical strength with higher accuracy.
